ETH Zurich's rotation-XMPI Captures 3D Metal Printing at 25,000 Frames per Second
A three-beam X-ray setup plus a self-supervised neural reconstruction hits 25,000 3D volumes per second, a 250x jump over prior operando tomography.
Takeaways
- Rotation-XMPI achieves 25,000 3D reconstructions per second, a 250x jump over prior operando LPBF tomography.
- Three synchronized X-ray beamlets plus slow 25 Hz rotation decouple temporal resolution from spin speed.
- Reconstruction uses an adapted X-Hexplane self-supervised deep-learning framework with tensor-factorized 4D representation.
- Demonstrated on alumina laser remelting at MAX IV's ForMAX beamline, resolving keyhole formation and collapse.
- Full 4D reconstruction runs in under an hour on a single RTX 4070 Ti with 10.5 GB VRAM.
- Code, datasets, and movies will be released upon peer-reviewed publication of the work.
Watching a laser weld metal or ceramic in 3D as it happens has been essentially impossible. The physics inside a melt pool, with tiny keyholes forming and collapsing, molten flow, and vapor pressure, all evolves on microsecond timescales, while the leading technique for peering inside opaque samples requires spinning them fast enough to capture a full angular sweep. A team from ETH Zurich, Lund University, Paul Scherrer Institut, and MAX IV has now broken that constraint with a method they call rotation-XMPI, delivering 4D reconstructions at 25,000 volumes per second.
Why melt pools stayed invisible in 3D
Laser powder bed fusion (LPBF), the dominant technique for metal and ceramic 3D printing, is governed by rapid melt-pool dynamics that need to be resolved in 3D plus time. Operando X-ray tomography is fundamentally constrained by how fast you can spin the sample, because each tomogram needs roughly a 180 degree sweep of projections, which couples temporal resolution directly to rotation speed.
Cranking up rotation is not a free lunch. Centrifugal forces distort the fragile dynamics you are trying to observe, and even the best synchrotron tomoscopy setups have capped LPBF imaging at around 100 volumes per second, with cutting-edge work reaching about 1,000. Melt-pool oscillations happen at 10 kHz and beyond, so conventional tomography smears them into unusable blur.
Three beamlets and a lazy spin
X-ray multi-projection imaging (XMPI) sidesteps the rotation problem by splitting one beam into several beamlets that hit the sample from different angles simultaneously, each captured by a dedicated detector. Crystals create the beamlets, separate detectors record them, and the result is sparse but instantaneous multi-view data that partially decouples temporal resolution from sample rotation. The catch: three or four views are nowhere near the hundreds of projections classical reconstruction algorithms expect.